chai1me/qwen2vl-receipt-lora

Qwen2-VL-2B Receipt Extraction (LoRA)

LoRA adapter fine-tuning Qwen2-VL-2B-Instruct to extract structured JSON (line items and total) from receipt images. Trained with QLoRA (4-bit) on a single Kaggle T4 GPU.

Results (60 held-out receipts, unseen during training)

Metric	Score
Valid JSON output	100%
Total exact-match	90%
Mean char-similarity on near-misses	0.957

Training

• Base: Qwen/Qwen2-VL-2B-Instruct (loaded 4-bit via unsloth/Qwen2-VL-2B-Instruct-bnb-4bit)
• Method: LoRA on ~0.8% of parameters (language + attention + MLP layers), vision encoder frozen
• Config: 2 epochs, effective batch size 8, learning rate 2e-4
• Data: CORD-v2, 400 train / 60 validation

Usage

from unsloth import FastVisionModel
model, tokenizer = FastVisionModel.from_pretrained("chai1me/qwen2vl-receipt-lora", load_in_4bit=True)
FastVisionModel.for_inference(model)

Limitations

Trained on café/restaurant receipts (CORD). Other document types need retraining. Field extraction only; not a forgery- or tamper-detection system.

Full code and evaluation: GitHub